transcribe phase A: the normalized result schema, its serializer, and the generated sample

[fyang0507]

Context

Decomposes #12, phase A of
TRANSCRIBE_IMPLEMENTATION_PLAN.md.

The schema is where fabrication becomes possible, so it comes first and its anti-fabrication
assertions are built with it rather than after. Everything downstream is shaped by it: a plan's
sample_output is this schema serialized from a placeholder, so plan cannot ship before it.

Scope

src/audio_cli/transcribe/result.py and sample.py.

{
  "schema_version": 1,
  "complete": true,
  "source": {"path": "...", "duration_seconds": 0.0, "timebase": "seconds"},
  "segments": [{"segment_id": "seg_0", "text": "...",
                "speaker": "...",          // iff diarization
                "start": 0.0, "end": 0.0,  // iff segment_timestamps
                "words": [{"word_id": "w_0", "text": "...", "start": 0.0, "end": 0.0}]}],
  "turns":              [],  // iff diarization
  "vad_regions":        [],  // iff vad
  "lid_regions":        [],  // iff lid
  "overlapped_speech":  [],  // iff overlapped_speech
  "abstentions":        [],  // floor artifact, always present
  "coverage":           {},  // iff complete is false
  "provenance": {"stack": "...", "outcomes": {}, "observed": {}, "plan": {}}
}

• One serializer. sample_output is a placeholder result through the same code path run
  uses. A second rendering path is the specific mistake to avoid: four stacks against nine
  capability names is ~2,048 combinations, and a hand-maintained example set would both rot and
  be infeasible.
• The gating rule is bidirectional. A key exists iff its capability was requested. Two
  capabilities license no key at all — languages and verbatim assert an interface and change
  no output shape — so their outcome is always produced and the key-set test must encode that
  they contribute nothing.
• Placeholder values. Timing and text are null, never 0.0 or a plausible string; 0.0 is
  a legal timestamp and would violate the no-synthesized-bounds floor the moment a consumer read
  it as measured. Enum fields show one legal member so a consumer can see the field is
  categorical. Real metadata the plan genuinely has — duration, path — is populated.
• coverage is over the source timeline: missing_intervals are the extents of units that
  did not run, covered_intervals their complement, covered_fraction covered over source
  duration, and covered_through_seconds the start of the first missing interval. That watermark
  is genuinely early when duration-bucketed ordering leaves the longest turns unprocessed
  (model_tests/benchmark/run_turn_attributed_mlx_asr.py:636), which is why the explicit
  interval lists sit beside it rather than instead of it.
• observed.total_wall_seconds is the sum of the stage walls, so the arithmetic
  tests/test_spec_docs.py already asserts stays checkable. A caller timing the command sees
  more, because interpreter startup and artifact writes are outside every stage — the same
  distinction model_tests/benchmark/run_interview_pipeline.py:201-206 draws.
• peak_rss_bytes is the maximum of the per-stage peaks, never their sum, and the per-stage
  numbers are kept so the claim is checkable rather than asserted.

Out of scope

The stack table, the planner, and any command surface — those are the next phase. No transport,
no adapters, no backend.

Acceptance

Assertions that must exist and must be shown to fail when violated.

• A key exists iff its capability was requested, asserted over every cell of VOCABULARY
  §Resolution's derivation table, including both refusal codes.
• "N/A" never appears as an attribution anywhere in a serialized document — walked over
  every mapping key, not only over segments, because turns carry speaker labels too.
• A placeholder document contains no 0.0 bound and no plausible placeholder string.
• languages and verbatim in the requested set add no key to the document.
• Constructing a document with a key for an unrequested capability fails the suite, and so
  does one with 0.0 where null belongs. An invariant that cannot fail is inert — the
  punctuation floor spent four review passes proving that.
• uv run --extra dev pytest green.

Depends on

#18 — the schema names complete and the abstention reason enum, which phase 0 adds to the documents.